Efficient and Flexible Method for Reducing Moderate-size Deep Neural Networks with Condensation